feat(orchestration): escalate repeated scope/manifest write-block to user approval (#301)

Small models frequently fail to comply with the WRITE_SCOPE/PATH_OUTSIDE_MANIFEST
remediation ("add its path via task_update affected_paths") and instead thrash the
same out-of-scope write call turn after turn, burning the session without ever
completing. After `tuning.escalateScopeAfterN` (default 3) consecutive rejections
of the SAME path for a scope/manifest reason, the orchestrator now reaches back to
the FIRST rejected invocation of that path (its pristine, pre-degraded arguments),
and routes it through the existing approval/pause flow instead of rejecting again.
On approval, a new WriteScopeGrantedEvent widens the effective write manifest for
that path for the rest of the session (mirroring how OutsidePathAccessGrantedEvent
already widens out-of-workspace reads) and the first attempt's write is executed.
On denial, the call is rejected as before. escalateScopeAfterN = 0 preserves the
old hard-block-forever behavior.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GeyGFXczJb8RUWGBKmkm6G

core/events/src/main/kotlin/com/correx/core/events/events/OrchestrationEvents.kt

@@ -53,6 +53,21 @@ data class OutsidePathAccessGrantedEvent(
     val path: String,
 ) : EventPayload
 
+/**
+ * Records that the operator approved widening the write scope/manifest to admit [path], after
+ * the same path was rejected `escalate_scope_after_n` times in a row (small models frequently
+ * fail to comply with the WRITE_SCOPE/PATH_OUTSIDE_MANIFEST remediation and thrash instead —
+ * see #301). Folded the same way [OutsidePathAccessGrantedEvent] widens out-of-workspace reads:
+ * subsequent writes to this path this session are admitted without re-prompting.
+ */
+@Serializable
+@SerialName("WriteScopeGranted")
+data class WriteScopeGrantedEvent(
+    val sessionId: SessionId,
+    val stageId: StageId,
+    val path: String,
+) : EventPayload
+
 @Serializable
 @SerialName("OrchestrationPaused")
 data class OrchestrationPausedEvent(

core/inference/src/main/kotlin/com/correx/core/inference/DefaultInferenceRouter.kt

@@ -2,6 +2,7 @@ package com.correx.core.inference
 
 import com.correx.core.events.types.ProviderId
 import com.correx.core.events.types.StageId
+import kotlinx.coroutines.delay
 import kotlinx.coroutines.sync.Mutex
 import kotlinx.coroutines.sync.withLock
 import kotlin.time.Duration
@@ -21,6 +22,12 @@ class DefaultInferenceRouter(
     private val strategy: RoutingStrategy,
     private val cacheTtl: Duration = 5.seconds,
     private val timeSource: TimeSource = TimeSource.Monotonic,
+    // A provider that briefly drops (crash + qa-stack restart) shouldn't collapse into a hard
+    // NoEligibleProvider abort — give it bounded time to come back before giving up. This only
+    // applies when the capability IS configured on some provider but that provider is currently
+    // unhealthy; a capability nobody ever declared fails immediately (see routeCapabilityCandidates).
+    private val unavailableRetryAttempts: Int = 3,
+    private val unavailableRetryDelay: Duration = 2.seconds,
 ) : InferenceRouter {
 
     private val cache = mutableMapOf<ProviderId, HealthEntry>()
@@ -45,12 +52,49 @@ class DefaultInferenceRouter(
         }
     }
 
+    private suspend fun refreshedHealth(provider: InferenceProvider): ProviderHealth =
+        lockFor(provider.id).withLock {
+            val fresh = provider.healthCheck()
+            mapMutex.withLock { cache[provider.id] = HealthEntry(fresh, timeSource.markNow()) }
+            fresh
+        }
+
+    // Bypasses healthCheck()/TTL entirely — writes Unavailable straight into the cache so the very
+    // next route() call gates on it, closing the ~18s reactive-poll lag (#300). A later cache-TTL
+    // expiry or the bounded-wait re-check in route() will naturally pick the provider back up once
+    // its own healthCheck() reports healthy again.
+    override suspend fun reportFailure(providerId: ProviderId, reason: String) {
+        lockFor(providerId).withLock {
+            mapMutex.withLock { cache[providerId] = HealthEntry(ProviderHealth.Unavailable(reason), timeSource.markNow()) }
+        }
+    }
+
     override suspend fun route(stageId: StageId, requiredCapabilities: Set<ModelCapability>): InferenceProvider {
         val candidates = requiredCapabilities
             .flatMap { registry.resolve(it) }
             .distinctBy { it.id }
             .ifEmpty { registry.listAll() }
-        val healthy = candidates.filter { cachedHealth(it) !is ProviderHealth.Unavailable }
+
+        // Nobody was ever configured with this capability set — no amount of waiting fixes that,
+        // fail fast instead of burning the bounded-wait budget below.
+        if (requiredCapabilities.isNotEmpty() &&
+            candidates.none { it.capabilities().map { c -> c.capability }.toSet().containsAll(requiredCapabilities) }
+        ) {
+            throw NoEligibleProviderException(stageId, requiredCapabilities)
+        }
+
+        var healthy = candidates.filter { cachedHealth(it) !is ProviderHealth.Unavailable }
+        var attempt = 0
+        while (healthy.isEmpty() && attempt < unavailableRetryAttempts) {
+            attempt++
+            log.warn(
+                "route: capability {} configured but all candidates unhealthy for stage={}" +
+                    " — waiting {} (attempt {}/{}) before declaring NoEligibleProvider",
+                requiredCapabilities, stageId.value, unavailableRetryDelay, attempt, unavailableRetryAttempts,
+            )
+            delay(unavailableRetryDelay)
+            healthy = candidates.filter { refreshedHealth(it) !is ProviderHealth.Unavailable }
+        }
         val selected = strategy.select(healthy, requiredCapabilities)
         // Post-selection re-check closes the TOCTOU window between initial filter and dispatch.
         when (val postHealth = selected.healthCheck()) {

core/inference/src/main/kotlin/com/correx/core/inference/InferenceRouter.kt

@@ -39,6 +39,14 @@ interface InferenceRouter {
         requiredCapabilities: Set<ModelCapability>,
         modelId: String?,
     ): InferenceProvider = route(stageId, requiredCapabilities)
+
+    /**
+     * Event-driven health gate: called the moment a connection-level failure is observed on
+     * [providerId] (e.g. mid-request connection drop), so the NEXT route() call sees it as
+     * unavailable immediately instead of waiting for the next periodic health poll/cache TTL to
+     * catch up. Default no-op for routers that don't cache health.
+     */
+    suspend fun reportFailure(providerId: com.correx.core.events.types.ProviderId, reason: String) = Unit
 }
 
 class NoEligibleProviderException(

core/kernel/src/main/kotlin/com/correx/core/kernel/orchestration/OrchestrationTuning.kt

@@ -41,4 +41,12 @@ data class OrchestrationTuning(
     val recoveryRouteBudget: Int = 2,
     /** Budget for tier-2 intent-holder arbiter re-routing. */
     val intentRouteBudget: Int = 2,
+    /**
+     * After this many consecutive WRITE_SCOPE/PATH_OUTSIDE_MANIFEST rejections of the SAME path in
+     * a session, stop hard-rejecting and escalate to user approval instead (#301) — small models
+     * frequently fail to comply with the remediation and thrash rather than widen scope themselves.
+     * 0 disables escalation (falls back to the hard-block-forever behavior). A headless session
+     * (no approver connected) auto-rejects the escalated prompt rather than hanging.
+     */
+    val escalateScopeAfterN: Int = 3,
 )

core/kernel/src/main/kotlin/com/correx/core/kernel/orchestration/SessionOrchestrator.kt

@@ -300,7 +300,15 @@ abstract class SessionOrchestrator(
         // ToolCalling score and routes the stage to the best tool-caller.
         val requiredCapabilities = stageConfig.requiredCapabilities +
             if (withTools && stageConfig.allowedTools.isNotEmpty()) setOf(ModelCapability.ToolCalling) else emptySet()
-        val provider = inferenceRouter.route(stageId, requiredCapabilities, stageConfig.modelId)
+        // Routing itself can fail transiently (provider mid-crash-recovery) — it must be retryable
+        // like any other inference failure, not escape and kill the whole session (see #299).
+        val provider = try {
+            inferenceRouter.route(stageId, requiredCapabilities, stageConfig.modelId)
+        } catch (e: CancellationException) {
+            throw e
+        } catch (e: Exception) {
+            return InferenceResult.Failed(e.message ?: "routing failed")
+        }
         log.debug(
             "[Orchestrator] inference session={} stage={} provider={} timeoutMs={}",
             sessionId.value, stageId.value, provider.id.value, timeoutMs,
@@ -401,6 +409,13 @@ abstract class SessionOrchestrator(
         } catch (e: CancellationException) {
             throw e // never swallow
         } catch (e: Exception) {
+            if (isConnectionLevelFailure(e)) {
+                // Mark it down NOW instead of waiting for the next periodic health poll (~18s lag,
+                // see #300) — the retry's route() call must see this provider as unavailable
+                // immediately so it gates/waits (#299) rather than instantly re-selecting the dead
+                // provider again.
+                inferenceRouter.reportFailure(provider.id, e.message ?: "connection failure")
+            }
             emit(
                 sessionId,
                 InferenceFailedEvent(
@@ -415,6 +430,19 @@ abstract class SessionOrchestrator(
         }
     }
 
+    // Connection-level failures (provider crashed/restarting mid-request) should gate routing
+    // immediately; other failures (bad request, model error, HTTP 4xx) should not mark the
+    // provider down since the provider itself is still reachable.
+    private fun isConnectionLevelFailure(e: Exception): Boolean {
+        val message = e.message.orEmpty()
+        return e is java.net.ConnectException ||
+            e is java.net.SocketException ||
+            e is java.io.IOException || // covers ktor/CIO's IOException, which extends java.io.IOException on the JVM
+            message.contains("prematurely closed", ignoreCase = true) ||
+            message.contains("connection refused", ignoreCase = true) ||
+            message.contains("connection reset", ignoreCase = true)
+    }
+
     // --- token estimation ---
 
     internal open suspend fun estimateTokens(content: String): Int {

core/kernel/src/main/kotlin/com/correx/core/kernel/orchestration/SessionOrchestratorToolExec.kt

@@ -31,6 +31,7 @@ import com.correx.core.events.events.ToolExecutionRejectedEvent
 import com.correx.core.events.events.ToolInvocationRequestedEvent
 import com.correx.core.events.events.ToolReceipt
 import com.correx.core.events.events.ToolRequest
+import com.correx.core.events.events.WriteScopeGrantedEvent
 import com.correx.core.events.risk.RiskAction
 import com.correx.core.events.risk.RiskSummary
 import com.correx.core.toolintent.ToolCallAssessmentInput
@@ -218,14 +219,36 @@ internal suspend fun SessionOrchestrator.dispatchToolCalls(
         // Per-task write scope: while a task is claimed, narrow the manifest to its affected
         // paths (recorded on the task) so the implementer can't write outside its unit of work.
         // Falls back to the stage's static manifest when nothing is claimed.
-        val effectiveManifest = taskClaimCoordinator?.activeScope(sessionId)?.takeIf { it.isNotEmpty() }
+        val escalatedScope = escalatedWriteScopePaths(sessionId)
+        val effectiveManifest = (
+            taskClaimCoordinator?.activeScope(sessionId)?.takeIf { it.isNotEmpty() }
                 ?: stageConfig.writeManifest
+            ) + escalatedScope
         val plane2Risk: RiskSummary? = runPlane2Assessment(
             sessionId, stageId, invocationId, toolCall.function.name, request, tool, effectives,
             effectiveManifest,
         )?.let { assessment ->
             if (assessment.recommendedAction == RiskAction.BLOCK) {
                 val rationale = assessment.rationale.joinToString("; ")
+                // #301: a write repeatedly rejected for being outside the claimed task's scope or
+                // the stage's manifest — never for anything else — escalates to user approval
+                // after N same-path rejections instead of hard-blocking forever. Small models
+                // routinely fail to comply with the "add its path via task_update" remediation and
+                // thrash the same call rather than widen scope themselves.
+                val escalationPath = (parameters["path"] as? String)
+                    ?.takeIf { isScopeBlockRationale(rationale) }
+                val escalateAfterN = tuning.escalateScopeAfterN
+                if (escalationPath != null && escalateAfterN > 0) {
+                    val priorRejections = priorScopeRejections(sessionId, escalationPath)
+                    if (priorRejections.size >= escalateAfterN) {
+                        val firstAttempt = priorRejections.first().first
+                        return@flatMap escalateWriteScopeBlock(
+                            sessionId, stageId, invocationId, toolCall, tier, escalationPath,
+                            firstAttempt, effectives, toolCallReasoning, approvalMode, assessment,
+                            fileWrittenSlots,
+                        )
+                    }
+                }
                 // On a bad-path block, point the model at the closest real file so it fixes the
                 // path instead of retrying the same wrong guess (deep package paths are easy to
                 // misremember). Only fires when we can name a concrete match from the repo map.
@@ -509,6 +532,180 @@ internal suspend fun SessionOrchestrator.dispatchToolCalls(
     }
 }
 
+private fun parametersToArgumentsJson(parameters: Map<String, Any>): String =
+    kotlinx.serialization.json.buildJsonObject {
+        parameters.forEach { (k, v) ->
+            when (v) {
+                is List<*> -> put(
+                    k,
+                    kotlinx.serialization.json.buildJsonArray {
+                        v.forEach { add(kotlinx.serialization.json.JsonPrimitive(it.toString())) }
+                    },
+                )
+                else -> put(k, kotlinx.serialization.json.JsonPrimitive(v.toString()))
+            }
+        }
+    }.toString()
+
+/**
+ * #301: the Nth same-path WRITE_SCOPE/PATH_OUTSIDE_MANIFEST rejection routes into the existing
+ * approval/pause flow instead of hard-blocking again. [firstAttempt] is the FIRST invocation that
+ * was rejected for this path/reason (reached back into via [priorScopeRejections]) — its pristine
+ * arguments are what gets previewed and, on approval, executed; later attempts this session may
+ * have degraded (e.g. the model giving up and calling `task_update(action="block")` instead), so
+ * replaying those would not fulfil the original intent.
+ */
+internal suspend fun SessionOrchestrator.escalateWriteScopeBlock(
+    sessionId: SessionId,
+    stageId: StageId,
+    invocationId: ToolInvocationId,
+    toolCall: ToolCallRequest,
+    tier: Tier,
+    path: String,
+    firstAttempt: ToolInvocationRequestedEvent,
+    effectives: RunEffectives,
+    toolCallReasoning: String?,
+    approvalMode: ApprovalMode,
+    plane2Risk: RiskSummary?,
+    fileWrittenSlots: List<TypedArtifactSlot>,
+): List<ContextEntry> {
+    val sourceId = toolCall.id ?: invocationId.value
+    val assistantEntry = ContextEntry(
+        id = ContextEntryId(UUID.randomUUID().toString()),
+        layer = ContextLayer.L2,
+        sourceType = "assistantToolCall",
+        sourceId = sourceId,
+        content = Json.encodeToString(ToolCallRequest.serializer(), toolCall),
+        tokenEstimate = estimateTokens(toolCall.function.arguments),
+        role = EntryRole.ASSISTANT,
+        reasoning = toolCallReasoning,
+    )
+    suspend fun rejected(reason: String): List<ContextEntry> {
+        blockTaskOnScopeRejection(sessionId, toolCall.function.name, reason)
+        emit(
+            sessionId,
+            ToolExecutionRejectedEvent(
+                invocationId = invocationId,
+                sessionId = sessionId,
+                toolName = toolCall.function.name,
+                tier = tier,
+                reason = reason,
+            ),
+        )
+        return listOf(
+            assistantEntry,
+            ContextEntry(
+                id = ContextEntryId(UUID.randomUUID().toString()),
+                layer = ContextLayer.L2,
+                sourceType = "toolResult",
+                sourceId = sourceId,
+                content = "BLOCKED: $reason",
+                tokenEstimate = estimateTokens(reason),
+                role = EntryRole.TOOL,
+            ),
+        )
+    }
+
+    val projectId = effectives.policy?.workspaceRoot?.let { ProjectIdentity.of(it.toString()) }
+    val approvalCtx = ApprovalContext(
+        identity = ApprovalScopeIdentity(sessionId, stageId, projectId = projectId),
+        mode = approvalMode,
+    )
+    val requestId = ApprovalRequestId(UUID.randomUUID().toString())
+    val toolPreview = computeToolPreview(
+        firstAttempt.toolName, firstAttempt.request.parameters, effectives.policy?.workspaceRoot,
+    )
+    val previewArguments = parametersToArgumentsJson(firstAttempt.request.parameters)
+    val domainRequest = DomainApprovalRequest(
+        id = requestId,
+        tier = firstAttempt.tier,
+        validationReportId = ValidationReportId(UUID.randomUUID().toString()),
+        riskSummaryId = null,
+        timestamp = Clock.System.now(),
+        toolName = firstAttempt.toolName,
+        preview = toolPreview ?: previewArguments.take(200),
+    )
+    val sessionGrants = approvalRepository.getApprovalState(sessionId).grants.values
+    val ledgerGrants = approvalRepository.getApprovalState(GRANT_LEDGER_SESSION_ID).grants.values
+    val activeGrants = (sessionGrants + ledgerGrants).toList()
+    val engineDecision = approvalEngine.evaluate(domainRequest, approvalCtx, activeGrants, Clock.System.now())
+    val approved: Boolean
+    val denyReason: String?
+    if (engineDecision.state == ApprovalStatus.COMPLETED) {
+        // Headless/no-approver sessions resolve here (e.g. approvalMode DENY auto-completes to a
+        // denial) — the escalation never hangs waiting on a human who isn't connected.
+        emitDecisionResolved(sessionId, domainRequest, engineDecision)
+        approved = engineDecision.isApproved
+        denyReason = engineDecision.reason
+    } else {
+        val deferred = CompletableDeferred<ApprovalDecision>()
+        pendingApprovals[requestId] = deferred
+        emit(sessionId, OrchestrationPausedEvent(sessionId, stageId, "APPROVAL_PENDING"))
+        emit(
+            sessionId,
+            ApprovalRequestedEvent(
+                requestId = requestId,
+                tier = firstAttempt.tier,
+                validationReportId = domainRequest.validationReportId,
+                riskSummaryId = null,
+                riskSummary = plane2Risk,
+                sessionId = sessionId,
+                stageId = stageId,
+                projectId = null,
+                toolName = firstAttempt.toolName,
+                preview = toolPreview ?: previewArguments.take(200),
+            ),
+        )
+        val userDecision = try {
+            deferred.await()
+        } finally {
+            pendingApprovals.remove(requestId)
+        }
+        emitDecisionResolved(sessionId, domainRequest, userDecision)
+        emit(sessionId, OrchestrationResumedEvent(sessionId, stageId))
+        approved = userDecision.isApproved
+        denyReason = userDecision.reason
+    }
+    if (!approved) {
+        return rejected(denyReason ?: "scope-widening request denied")
+    }
+
+    // Approved: widen the write scope for this path for the rest of the session (future writes to
+    // it skip straight past the risk plane, mirroring OutsidePathAccessGrantedEvent), then execute
+    // the FIRST rejected attempt's pristine write rather than the model's current — possibly
+    // degraded — call.
+    emit(sessionId, WriteScopeGrantedEvent(sessionId, stageId, path))
+    val executor = effectives.executor ?: return rejected("no executor available to apply the approved write")
+    val tool = effectives.registry?.resolve(firstAttempt.toolName)
+    val result = executor.execute(firstAttempt.request)
+    val rendered = renderToolResult(firstAttempt.toolName, tool, result)
+    recordToolExecution(
+        sessionId,
+        stageId,
+        toolCall.copy(function = toolCall.function.copy(name = firstAttempt.toolName)),
+        invocationId,
+        tier,
+        result,
+        tool as? FileAffectingTool,
+        firstAttempt.request,
+        fileWrittenSlots,
+        rendered.fullOutputHash,
+    )
+    return listOf(
+        assistantEntry,
+        ContextEntry(
+            id = ContextEntryId(UUID.randomUUID().toString()),
+            layer = ContextLayer.L2,
+            sourceType = "toolResult",
+            sourceId = sourceId,
+            content = "APPROVED: the user widened write scope to admit '$path' after repeated rejection; " +
+                "the original write is applied. ${rendered.content}",
+            tokenEstimate = estimateTokens(rendered.content),
+            role = EntryRole.TOOL,
+        ),
+    )
+}
+
 /**
  * A rejected [SCOPE_PROPOSAL_TOOL] call means the operator denied widening the claimed task's
  * scope — block the task so the loop advances instead of the implementer retrying the same

core/kernel/src/main/kotlin/com/correx/core/kernel/orchestration/SessionOrchestratorWorkspace.kt

@@ -4,7 +4,9 @@ import com.correx.core.events.events.ToolCallAssessedEvent
 import com.correx.core.events.events.RepoMapComputedEvent
 import com.correx.core.events.events.OutsidePathAccessGrantedEvent
 import com.correx.core.events.events.ToolExecutionCompletedEvent
+import com.correx.core.events.events.ToolExecutionRejectedEvent
 import com.correx.core.events.events.ToolInvocationRequestedEvent
+import com.correx.core.events.events.WriteScopeGrantedEvent
 import com.correx.core.events.risk.RiskAction
 import com.correx.core.toolintent.SessionContextProjection
 import com.correx.core.events.events.TransitionExecutedEvent
@@ -68,6 +70,55 @@ internal fun SessionOrchestrator.grantedOutsidePaths(sessionId: SessionId): Set<
         .map { it.path }
         .toSet()
 
+/**
+ * Write-scope/manifest paths the operator has approved widening into this session (folded from
+ * [WriteScopeGrantedEvent] — see #301). Replay-safe: reads the log, never re-prompts a path once
+ * granted.
+ */
+
+internal fun SessionOrchestrator.escalatedWriteScopePaths(sessionId: SessionId): Set<String> =
+    eventStore.read(sessionId)
+        .mapNotNull { it.payload as? WriteScopeGrantedEvent }
+        .filter { it.sessionId == sessionId }
+        .map { it.path }
+        .toSet()
+
+/** The rule codes whose BLOCK rationale (`"[CODE] message"`, see [toRiskSummary]) makes a
+ * rejection eligible for the #301 same-path escalation — a write rejected purely for being
+ * outside the claimed task's scope or the stage's manifest, not for any other reason (path
+ * traversal, privileged location, etc. are never escalated). */
+private val SCOPE_BLOCK_CODES = setOf("WRITE_SCOPE", "PATH_OUTSIDE_MANIFEST")
+
+internal fun SessionOrchestrator.isScopeBlockRationale(rationale: String): Boolean =
+    SCOPE_BLOCK_CODES.any { rationale.contains("[$it]") }
+
+/**
+ * Every (first-attempt-invocation, rejection) pair this session where [path] was rejected for a
+ * WRITE_SCOPE/PATH_OUTSIDE_MANIFEST reason, oldest first. Reached back into purely by folding the
+ * event log — no branching/replay-engine change needed (#301). The invocation carries the
+ * pristine [ToolRequest] parameters from that attempt, which may differ from the model's current
+ * (possibly degraded) call.
+ */
+
+internal fun SessionOrchestrator.priorScopeRejections(
+    sessionId: SessionId,
+    path: String,
+): List<Pair<ToolInvocationRequestedEvent, ToolExecutionRejectedEvent>> {
+    val events = eventStore.read(sessionId)
+    val invocationsById = events
+        .mapNotNull { it.payload as? ToolInvocationRequestedEvent }
+        .filter { it.sessionId == sessionId }
+        .associateBy { it.invocationId }
+    return events
+        .mapNotNull { it.payload as? ToolExecutionRejectedEvent }
+        .filter { it.sessionId == sessionId && isScopeBlockRationale(it.reason) }
+        .mapNotNull { rejected ->
+            val invocation = invocationsById[rejected.invocationId] ?: return@mapNotNull null
+            val invocationPath = invocation.request.parameters["path"] as? String
+            if (invocationPath == path) invocation to rejected else null
+        }
+}
+
 /**
  * Returns true when the agent must be offered only read-only tools on the next inference turn.
  * Active from the moment a READ_BEFORE_WRITE block lands until a FILE_READ completion follows it.

examples/workflows/prompts/analyst_freestyle.md

@@ -6,16 +6,27 @@ Before deriving requirements, check for existing work: `task_search` for related
 blocking tasks and `task_context` to load any the goal names. Fold what you find into the
 analysis rather than re-deriving it; flag a duplicate instead of restating it.
 
-Then frame the work as a task (per the task policy):
+Then frame the work as a task (per the task policy). A run always names exactly ONE task — never
-- If a task already covers this work, name its id (e.g. `auth-142`) in the analysis.
+a parent/epic — as the thing it will work:
-- If the goal is a single coherent unit one run can carry to review, `task_create` one and name its
+- If an existing **leaf** task already covers this work (no children of its own), name its id
-  id.
+  (e.g. `auth-142`) in the analysis.
-- If the goal has **dependency seams** (a thing that must land before another) or **independent
+- If an existing task covering this goal is itself a **parent/epic** (has `DEPENDS_ON`-linked
-  review/handoff points** (a piece worth shipping or reviewing on its own), `task_decompose` it into
+  children, whether from a past run's `task_decompose` or found via `task_search`/`task_context`),
-  a parent epic + `DEPENDS_ON`-linked children — one approval for the whole graph. A session works
+  do **not** name the epic and do not decompose it again. Name the single ready child instead — the
-  one task at a time, so the children are claimed by *later* runs as they unblock; don't over-split.
+  one with no unmet dependency. If every child is already blocked/claimed, name the closest-to-ready
-- After decomposing, **name in the analysis the single task this run will work** — the one already
+  one and note in the analysis that this run is unblocking it, not completing the epic.
-  ready (no unmet dependency, e.g. the scaffold). Leave the blocked siblings for future runs.
+- If no task covers this work yet and the goal is a single coherent unit one run can carry to
+  review, `task_create` one and name its id.
+- If no task covers this work yet and the goal has **dependency seams** (a thing that must land
+  before another) or **independent review/handoff points** (a piece worth shipping or reviewing on
+  its own), `task_decompose` it into a parent epic + `DEPENDS_ON`-linked children — one approval for
+  the whole graph. A session works one task at a time, so the children are claimed by *later* runs
+  as they unblock; don't over-split. Then name the single ready child (the one already unblocked,
+  e.g. the scaffold) — never the epic itself.
+
+There is always exactly one task id to name by the time you call `emit_artifact` — if you find
+yourself unsure whether to name a parent or a child, the answer is always the child. Do not loop on
+this decision.
 
 Either way later stages thread the named task through the plan; the rest wait to be claimed.
 

testing/contracts/src/test/kotlin/com/correx/testing/contracts/model/inference/DefaultInferenceRouterTest.kt

@@ -13,8 +13,10 @@ import com.correx.core.inference.RoutingStrategy
 import com.correx.testing.fixtures.inference.MockInferenceProvider
 import kotlinx.coroutines.runBlocking
 import org.junit.jupiter.api.Assertions.assertSame
+import org.junit.jupiter.api.Assertions.assertTrue
 import org.junit.jupiter.api.Test
 import org.junit.jupiter.api.assertThrows
+import kotlin.time.Duration.Companion.milliseconds
 
 class DefaultInferenceRouterTest {
 
@@ -157,4 +159,82 @@ class DefaultInferenceRouterTest {
         val result = router.route(stage, setOf(ModelCapability.General), "llama-cpp:sick")
         assertSame(healthy, result)
     }
+
+    // ── bounded wait for a briefly-absent provider (#299) ─────────────────────
+
+    @Test
+    fun `waits for the sole capable provider to recover instead of failing immediately`(): Unit = runBlocking {
+        var healthChecks = 0
+        val recovering = object : InferenceProvider by provider("a", ModelCapability.ToolCalling) {
+            override suspend fun healthCheck(): ProviderHealth {
+                healthChecks++
+                return if (healthChecks < 3) ProviderHealth.Unavailable("connection dropped") else ProviderHealth.Healthy
+            }
+        }
+        val router = DefaultInferenceRouter(
+            registryOf(recovering),
+            firstStrategy(),
+            unavailableRetryAttempts = 5,
+            unavailableRetryDelay = 5.milliseconds,
+        )
+        val result = router.route(stage, setOf(ModelCapability.ToolCalling))
+        assertSame(recovering, result)
+        assertTrue(healthChecks >= 3) { "expected at least 3 health checks, got $healthChecks" }
+    }
+
+    @Test
+    fun `still throws NoEligibleProviderException if the sole provider never recovers within the bound`() {
+        val neverRecovers = MockInferenceProvider(
+            id = ProviderId("a"),
+            declaredCapabilities = setOf(CapabilityScore(ModelCapability.ToolCalling, 1.0)),
+            health = ProviderHealth.Unavailable("still down"),
+        )
+        val router = DefaultInferenceRouter(
+            registryOf(neverRecovers),
+            throwingStrategy(),
+            unavailableRetryAttempts = 2,
+            unavailableRetryDelay = 5.milliseconds,
+        )
+        assertThrows<NoEligibleProviderException> {
+            runBlocking { router.route(stage, setOf(ModelCapability.ToolCalling)) }
+        }
+    }
+
+    @Test
+    fun `fails fast without waiting when the capability was never configured on any provider`(): Unit = runBlocking {
+        val p = provider("a", ModelCapability.General) // does not declare ToolCalling
+        val router = DefaultInferenceRouter(
+            registryOf(p),
+            throwingStrategy(),
+            unavailableRetryAttempts = 5,
+            unavailableRetryDelay = 10_000.milliseconds, // would time the test out if the wait loop ran
+        )
+        assertThrows<NoEligibleProviderException> {
+            router.route(stage, setOf(ModelCapability.ToolCalling))
+        }
+    }
+
+    // ── event-driven health gating (#300) ─────────────────────────────────────
+
+    @Test
+    fun `reportFailure gates the very next route call without waiting for a health poll`(): Unit = runBlocking {
+        var healthCheckCount = 0
+        val alwaysClaimsHealthy = object : InferenceProvider by provider("a", ModelCapability.General) {
+            override suspend fun healthCheck(): ProviderHealth {
+                healthCheckCount++
+                return ProviderHealth.Healthy // the provider's own health probe hasn't caught up yet
+            }
+        }
+        val backup = provider("b", ModelCapability.General)
+        val router = DefaultInferenceRouter(registryOf(alwaysClaimsHealthy, backup), firstStrategy())
+
+        // Sanity: before reportFailure, routes to the first (still "healthy") provider.
+        assertSame(alwaysClaimsHealthy, router.route(stage, setOf(ModelCapability.General)))
+
+        router.reportFailure(alwaysClaimsHealthy.id, "connection reset")
+
+        // Immediately after — no delay, no waiting for the next poll — routing must avoid it.
+        val result = router.route(stage, setOf(ModelCapability.General))
+        assertSame(backup, result)
+    }
 }

testing/integration/src/test/kotlin/SessionOrchestratorIntegrationTest.kt

@@ -43,10 +43,14 @@ import com.correx.core.inference.InferenceProvider
 import com.correx.core.inference.InferenceRepository
 import com.correx.core.inference.InferenceRequest
 import com.correx.core.inference.InferenceResponse
+import com.correx.core.inference.DefaultInferenceRouter
 import com.correx.core.inference.InferenceRouter
 import com.correx.core.inference.InferenceState
 import com.correx.core.inference.ModelCapability
+import com.correx.core.inference.NoEligibleProviderException
 import com.correx.core.inference.ProviderHealth
+import com.correx.core.inference.ProviderRegistry
+import com.correx.core.inference.RoutingStrategy
 import com.correx.core.inference.Token
 import com.correx.core.inference.TokenUsage
 import com.correx.core.inference.Tokenizer
@@ -216,6 +220,84 @@ class SessionOrchestratorIntegrationTest {
         assertTrue(failed.retryExhausted)
     }
 
+    @Test
+    fun `transient NoEligibleProviderException on a retry is tolerated, not a session-killing crash (#299)`(): Unit =
+        runBlocking {
+            val sessionId = SessionId("s2b")
+            val config = OrchestrationConfig(
+                retryPolicy = RetryPolicy(maxAttempts = 3, backoffMs = 0),
+            )
+            val graph = threeStageGraph()
+            var routeCalls = 0
+
+            val recoveringOrchestrator = DefaultSessionOrchestrator(
+                repositories = repositories,
+                engines = engines.copy(
+                    inferenceRouter = object : InferenceRouter {
+                        override suspend fun route(stageId: StageId, requiredCapabilities: Set<ModelCapability>): InferenceProvider {
+                            routeCalls++
+                            // First attempt hits the provider mid-crash — subsequent retries find it back.
+                            if (routeCalls == 1) throw NoEligibleProviderException(stageId, requiredCapabilities)
+                            return MockInferenceProvider()
+                        }
+                    },
+                ),
+                retryCoordinator = retryCoordinator,
+                artifactStore = artifactStore,
+                decisionJournalRepository = decisionJournalRepository,
+            )
+            recoveringOrchestrator.run(sessionId, graph, config)
+
+            val events = eventStore.read(sessionId)
+            // A crash would have skipped straight to an unhandled failure with no WorkflowCompletedEvent
+            // and no orderly retry — assert the session survived and made it through via the retry path.
+            assertTrue(routeCalls > 1) { "expected routing to be retried, only called $routeCalls time(s)" }
+            assertNotNull(events.find { it.payload is WorkflowCompletedEvent })
+            assertTrue(events.none { it.payload is WorkflowFailedEvent })
+        }
+
+    @Test
+    fun `connection-level inference failure gates routing away from the dead provider immediately (#300)`(): Unit =
+        runBlocking {
+            val sessionId = SessionId("s2c")
+            val config = OrchestrationConfig(
+                retryPolicy = RetryPolicy(maxAttempts = 3, backoffMs = 0),
+            )
+            val graph = threeStageGraph()
+
+            val dead = MockInferenceProvider(
+                id = ProviderId("dead"),
+                forcedFailure = "Failed to parse HTTP response: the server prematurely closed the connection",
+            )
+            val backup = MockInferenceProvider(id = ProviderId("backup"))
+            // Router's own health probe (dead.healthCheck()) still reports Healthy — it hasn't
+            // polled yet — so only the event-driven reportFailure() from the orchestrator's catch
+            // block (not the periodic poll) can gate the retry away from `dead`.
+            val registry = object : ProviderRegistry {
+                override fun register(provider: com.correx.core.inference.InferenceProvider) = Unit
+                override fun resolve(capability: ModelCapability) = listOf(dead, backup)
+                override fun listAll() = listOf(dead, backup)
+                override suspend fun healthCheckAll() = mapOf(dead.id to ProviderHealth.Healthy, backup.id to ProviderHealth.Healthy)
+            }
+            val router = DefaultInferenceRouter(
+                registry = registry,
+                strategy = RoutingStrategy { candidates, _ -> candidates.first() },
+            )
+
+            val recoveringOrchestrator = DefaultSessionOrchestrator(
+                repositories = repositories,
+                engines = engines.copy(inferenceRouter = router),
+                retryCoordinator = retryCoordinator,
+                artifactStore = artifactStore,
+                decisionJournalRepository = decisionJournalRepository,
+            )
+            recoveringOrchestrator.run(sessionId, graph, config)
+
+            val events = eventStore.read(sessionId)
+            assertNotNull(events.find { it.payload is WorkflowCompletedEvent })
+            assertEquals(1, dead.inferCallCount) { "dead provider should only be attempted once, then gated out" }
+        }
+
     @Test
     fun `stage that matches no transition retries through the budget before failing`(): Unit = runBlocking {
         val sessionId = SessionId("s-no-transition-retry")

testing/integration/src/test/kotlin/ToolCallGateTest.kt

@@ -18,6 +18,8 @@ import com.correx.core.events.events.ToolExecutionCompletedEvent
 import com.correx.core.events.events.ToolExecutionRejectedEvent
 import com.correx.core.events.events.ToolInvocationRequestedEvent
 import com.correx.core.events.events.ToolRequest
+import com.correx.core.events.events.WriteScopeGrantedEvent
+import com.correx.core.toolintent.rules.ManifestContainmentRule
 import com.correx.core.toolintent.rules.ReadBeforeWriteRule
 import com.correx.core.toolintent.rules.ReferenceExistsRule
 import com.correx.core.tools.contract.ParamRole
@@ -50,6 +52,7 @@ import com.correx.core.kernel.orchestration.OrchestrationConfig
 import com.correx.core.kernel.orchestration.OrchestrationProjector
 import com.correx.core.kernel.orchestration.OrchestrationRepository
 import com.correx.core.kernel.orchestration.OrchestratorEngines
+import com.correx.core.kernel.orchestration.OrchestrationTuning
 import com.correx.core.kernel.orchestration.OrchestratorRepositories
 import com.correx.core.kernel.retry.DefaultRetryCoordinator
 import com.correx.core.risk.DefaultRiskAssessor
@@ -717,6 +720,178 @@ class ToolCallGateTest {
         )
     }
 
+    /**
+     * #301: a write repeatedly rejected for being outside the stage's declared manifest escalates
+     * to user approval after N same-path rejections instead of hard-blocking forever. On approval
+     * the FIRST attempt's pristine write is applied (widening the manifest for the session), and
+     * later attempts to the same path skip the prompt entirely.
+     */
+    @Test
+    fun `Nth same-path PATH_OUTSIDE_MANIFEST rejection escalates to approval, which then executes the first attempt (#301)`(): Unit =
+        runBlocking {
+            val blockedPath = "/work/scratch/blocked.kt"
+            // T2 so the approval engine cannot auto-approve under PROMPT mode (T1 would auto-approve
+            // and the escalation would never actually pause) — mirrors the other PROMPT_USER tests.
+            val fileWriteTool = object : Tool {
+                override val name = "file_write"
+                override val description = "write"
+                override val parametersSchema: JsonObject = buildJsonObject {}
+                override val tier = Tier.T2
+                override val requiredCapabilities: Set<ToolCapability> = setOf(ToolCapability.FILE_WRITE)
+                override val paramRoles: Map<String, ParamRole> = mapOf("path" to ParamRole.PATH)
+                override fun validateRequest(request: ToolRequest): ValidationResult = ValidationResult.Valid
+            }
+            val toolRegistry = object : ToolRegistry {
+                override fun resolve(name: String): Tool? = if (name == "file_write") fileWriteTool else null
+                override fun all(): List<Tool> = listOf(fileWriteTool)
+            }
+
+            var turnCount = 0
+            val provider = object : InferenceProvider {
+                override val id = ProviderId("scope-escalate")
+                override val name = "scope-escalate"
+                override val tokenizer: Tokenizer = MockTokenizer()
+                val requests = mutableListOf<InferenceRequest>()
+                override suspend fun infer(request: InferenceRequest): InferenceResponse {
+                    requests += request
+                    turnCount++
+                    // Turns 1-3 keep retrying the exact same out-of-manifest write; turn 4 (after
+                    // the escalated approval resolves) stops so the run completes.
+                    return if (turnCount <= 3) {
+                        InferenceResponse(
+                            request.requestId, "", FinishReason.ToolCall, TokenUsage(1, 1), 0,
+                            listOf(
+                                ToolCallRequest(
+                                    id = "tc-$turnCount",
+                                    function = ToolCallFunction("file_write", """{"path":"$blockedPath"}"""),
+                                ),
+                            ),
+                        )
+                    } else {
+                        InferenceResponse(request.requestId, "done", FinishReason.Stop, TokenUsage(1, 1), 0)
+                    }
+                }
+                override suspend fun healthCheck(): ProviderHealth = ProviderHealth.Healthy
+                override fun capabilities(): Set<CapabilityScore> = setOf(CapabilityScore(ModelCapability.General, 1.0))
+            }
+
+            val fakeProbe = object : com.correx.core.toolintent.WorldProbe {
+                override fun exists(path: Path): Boolean = true
+                override fun resolveReal(path: Path): Path = path.toAbsolutePath().normalize()
+            }
+
+            val eventStore = InMemoryEventStore()
+            val artifactStore = NoopArtifactStore()
+            val assessor = ToolCallAssessor(listOf(ManifestContainmentRule()))
+            val policy = WorkspacePolicy(workspace)
+            val executor = RecordingExecutor()
+
+            val repositories = OrchestratorRepositories(
+                eventStore = eventStore,
+                inferenceRepository = InferenceRepository(object : EventReplayer<InferenceState> {
+                    override fun rebuild(sessionId: SessionId) = InferenceState()
+                }),
+                orchestrationRepository = OrchestrationRepository(
+                    DefaultEventReplayer(eventStore, OrchestrationProjector(DefaultOrchestrationReducer())),
+                ),
+                sessionRepository = DefaultSessionRepository(
+                    DefaultEventReplayer(eventStore, SessionProjector(DefaultSessionReducer())),
+                ),
+                artifactRepository = LiveArtifactRepository(eventStore, DefaultArtifactReducer()),
+                approvalRepository = DefaultApprovalRepository(
+                    DefaultEventReplayer(eventStore, ApprovalProjector(DefaultApprovalReducer())),
+                ),
+            )
+            val engines = OrchestratorEngines(
+                transitionResolver = DefaultTransitionResolver { _, _ -> true },
+                contextPackBuilder = ContextFixtures.simpleBuilder(),
+                inferenceRouter = object : InferenceRouter {
+                    override suspend fun route(stageId: StageId, requiredCapabilities: Set<ModelCapability>): InferenceProvider = provider
+                },
+                validationPipeline = ValidationPipeline(validators = listOf(cyclePolicyMissingValidator())),
+                approvalEngine = DefaultApprovalEngine(),
+                riskAssessor = DefaultRiskAssessor(),
+                toolExecutor = executor,
+                toolRegistry = toolRegistry,
+                toolCallAssessor = assessor,
+                workspacePolicy = policy,
+                worldProbe = fakeProbe,
+            )
+            val orchestrator = DefaultSessionOrchestrator(
+                repositories = repositories,
+                engines = engines,
+                retryCoordinator = DefaultRetryCoordinator(eventStore),
+                artifactStore = artifactStore,
+                decisionJournalRepository = DefaultDecisionJournalRepository(
+                    DefaultEventReplayer(eventStore, DecisionJournalProjector(DefaultDecisionJournalReducer())),
+                ),
+                // Escalate after 2 same-path rejections instead of the default 3, so the test's
+                // 3rd attempt is the one that triggers the approval prompt.
+                tuning = OrchestrationTuning(escalateScopeAfterN = 2),
+            )
+
+            val sessionId = SessionId("scope-escalate")
+            val graph = WorkflowGraph(
+                id = "scope-escalate-test",
+                stages = mapOf(
+                    StageId("A") to StageConfig(
+                        allowedTools = setOf("file_write"),
+                        writeManifest = listOf("allowed/**"),
+                    ),
+                ),
+                transitions = setOf(
+                    TransitionEdge(TransitionId("t1"), StageId("A"), StageId("done"), condition = { true }),
+                ),
+                start = StageId("A"),
+            )
+            val config = OrchestrationConfig(retryPolicy = RetryPolicy(maxAttempts = 1, backoffMs = 0))
+
+            val job = launch { orchestrator.run(sessionId, graph, config) }
+
+            // Wait for the 3rd attempt's escalation to raise an ApprovalRequestedEvent.
+            val approval = withTimeout(5_000) {
+                var req: ApprovalRequestedEvent? = null
+                while (req == null) {
+                    req = eventStore.read(sessionId).firstNotNullOfOrNull { it.payload as? ApprovalRequestedEvent }
+                    if (req == null) yield()
+                }
+                req
+            }
+            assertEquals("file_write", approval.toolName)
+
+            // Exactly 2 hard rejections happened before the escalation (turns 1 and 2); the 3rd
+            // attempt paused for approval instead of rejecting again.
+            val rejectionsBeforeApproval = eventStore.read(sessionId).count { it.payload is ToolExecutionRejectedEvent }
+            assertEquals(2, rejectionsBeforeApproval, "expected exactly 2 hard rejections before escalation")
+            assertTrue(!executor.executeCalled.get(), "executor must not run before the escalated approval resolves")
+
+            orchestrator.submitApprovalDecision(
+                approval.requestId,
+                ApprovalDecision(
+                    id = null,
+                    requestId = approval.requestId,
+                    outcome = ApprovalOutcome.APPROVED,
+                    state = ApprovalStatus.COMPLETED,
+                    tier = Tier.T2,
+                    contextSnapshot = ApprovalContext(
+                        identity = ApprovalScopeIdentity(sessionId, StageId("A"), projectId = null),
+                        mode = ApprovalMode.PROMPT,
+                    ),
+                    resolutionTimestamp = Clock.System.now(),
+                    reason = "approved widening",
+                ),
+            )
+
+            withTimeout(5_000) {
+                while (eventStore.read(sessionId).none { it.payload is WriteScopeGrantedEvent }) yield()
+            }
+            job.join()
+
+            assertTrue(executor.executeCalled.get(), "the approved write must execute")
+            val grant = eventStore.read(sessionId).firstNotNullOfOrNull { it.payload as? WriteScopeGrantedEvent }
+            assertEquals(blockedPath, grant?.path)
+        }
+
     @Test
     fun `empty allowedTools denies all domain tools and emits ToolExecutionRejectedEvent`(): Unit = runBlocking {
         val executor = RecordingExecutor()